The cycle’s argument is that seeing the machine clearly requires seeing through the narrative of inevitability that surrounds it. The ladder narrative of AI progress—from early computation to neural networks to AGI, each step the natural preparation for the next—functions precisely as the ladder of biological progress did in Gould’s time: it makes the present look inevitable and the future look predetermined, which relieves the humans who are actually making the choices of any sense that their choices matter. The thought experiment of replaying the tape dissolves this inevitability at its foundation. The specific AI that exists today is not the necessary product of any technological law. It is the contingent product of specific funding decisions, specific publishing events, specific institutional cultures, and specific economic conditions, any one of which might have been otherwise and would have produced a different branch.
The implication is not that the current trajectory is wrong but that it is chosen—and that different choices would produce different trajectories. This makes individual and institutional decisions at the present branch points consequential in a way the ladder narrative cannot accommodate. Regulation that shapes which approaches receive funding is a form of ecological selection pressure that will prune some body plans and favor others. Educational choices about what computational skills the next generation develops are a form of developmental constraint that will determine which variations are available when the next punctuation arrives. The tape is still running. The organisms alive at this moment—the transformer-based language models, the humans who use them, the institutions that deploy them—are not the inevitable products of technological law. They are the contingent inhabitants of a specific moment, and the specific future that emerges from this moment depends on the specific choices being made now.
The thought experiment also serves as a corrective to the adaptationist fallacy as applied to existing AI systems. The adaptationist reads every feature as a designed solution to a design problem. The Gouldian asks what the feature’s actual history was, and the answer is often surprising: capabilities celebrated as triumphs of AI design are frequently spandrels—unintended byproducts of an architecture designed for something else—and limitations treated as inherent to the category are frequently the accidents of the specific lineage, not features of all possible AI. Replaying the tape is the tool for asking which features of current systems are genuinely inherent to any sufficiently capable AI, and which are the contingent marks of this specific evolutionary path.
Gould developed the thought experiment in Wonderful Life: The Burgess Shale and the Nature of History (1989), a book-length argument built around the reinterpretation of the Burgess Shale fauna by Harry Whittington and his students in the 1970s. The Burgess Shale, a 508-million-year-old deposit in British Columbia, contains fossils of extraordinary diversity—animals that preserved soft parts, revealing body plans with no obvious modern descendants. Earlier interpretations had tried to fit the Burgess fauna into the ancestral lineages of modern phyla. Whittington’s team found that many of the organisms were too strange to fit any modern lineage: they were representatives of body plans that went extinct without issue, experiments of the Cambrian explosion that did not survive into the modern world.
Gould seized on this reinterpretation as evidence for his thesis about contingency. If the Cambrian explosion produced dozens of viable body plans and most went extinct—not because they were defective but because the specific contingent events of the subsequent half-billion years favored certain lineages—then the specific organisms that dominate the modern world are contingent survivors, not the predetermined products of any evolutionary law. The lineage that produced vertebrates survived, but there was nothing necessary about its survival. The tape replayed might favor a different Cambrian lineage, and the entire subsequent history of complex life would unfold differently.
The thought experiment attracted sharp criticism from Simon Conway Morris, who used the same Burgess Shale material to argue for convergent evolution: that the functional demands of complex life channel evolution toward similar forms regardless of historical path, so that intelligence, sensory systems, and locomotion would evolve again in any replay. The debate between Gould’s contingency and Conway Morris’s convergence is the deepest open question about the nature of evolution, and its analog in AI—whether any sufficiently capable AI would converge on something like the current architectures, or whether the specific transformer is a contingent survivor of a specific path—is equally unresolved and equally consequential.
Contingency, not inevitability. The specific outcome of any branching, historical process depends on the specific sequence of contingent events that shaped it, and alternative sequences would have produced alternative outcomes. The thought experiment operationalizes this claim: if you cannot say with confidence what a replay would produce, you cannot claim the present outcome was inevitable. The AI discourse routinely makes the inevitability claim; replaying the tape is the tool for testing it.
What the ladder conceals. The ladder narrative prunes the branches that did not survive and presents only the trunk, making the trunk look inevitable. The neural network winters, the extinction of the LISP machines, the two decades of suppressed connectionist research—each is a branch the ladder conceals. Gouldian contingency demands that these branches be restored to view, not as historical curiosities but as evidence that the path taken was chosen, not decreed. The alternatives that were suppressed remain alternatives that different choices might recover.
Contingency and responsibility. The most important implication of the thought experiment is about agency. If the trajectory is determined, individual choices are decorative. If the trajectory is contingent—if a different sequence of choices would produce a different AI landscape—then the choices being made now by the specific humans positioned at the current branch points carry genuine weight. Regulation, funding, education, and institutional design are all forms of selection pressure on the AI bush, and their effects are path-dependent: what they produce depends not only on the pressure but on the variation available at the moment of application.
The convergence counter-argument. The strongest objection to the thought experiment’s application to AI is the convergence argument: that the functional demands of language-processing and general reasoning are so strong that some lineage of AI would converge on something like the current systems regardless of path. If so, the specific contingencies—the 2017 paper, the neural network winter—are local details that shaped the timing but not the destination. Gould’s response would be that convergent evolution produces analogous features in different substrates, not identical organisms: eyes evolved many times, but no two lineages’ eyes are the same. Similarly, some form of capable AI might emerge from any path, but the specific capabilities, the specific failure modes, and the specific social consequences of the specific AI we have are products of the specific path, not predictions from any functional necessity.